Dialectical Claim Verification
with Source Provenance

2.1 Speed Asymmetry

False claims propagate at network speed. A 2018 MIT study published in Science found that false news reaches 1,500 people six times faster than true stories. Current corrective mechanisms operate on fundamentally incompatible timescales: Community Notes has a median response time exceeding 14 hours, and 90% of submitted notes never reach publication. Professional fact-checkers (PolitiFact, Snopes) produce high-quality analysis but cover a fraction of viral content, often days after peak virality.

2.2 Granularity Gap

Existing automated tools operate at the wrong level of abstraction. NewsGuard and Provenance rate publications, not individual articles or claims. A credible outlet can publish a misleading article; a low-credibility source can occasionally report accurately. Ground News visualizes media bias distribution but does not assess whether specific claims are true. Grok generates confident natural-language responses but fabricates citations: a Stanford Internet Observatory analysis found a 94% citation error rate in Grok's fact-checking responses.

2.3 Opacity of Incentives

No existing tool systematically connects factual claims to the financial incentives behind them. When a news outlet reports on legislation, users cannot easily determine that the outlet's parent company donated $2.3M to the bill's sponsors, or that the outlet's owner also owns companies affected by the legislation. This incentive layer is critical context for evaluating information but is entirely absent from current fact-checking tools.

4.1 Claim Decomposition

Input content (tweet text, image transcriptions, linked article text) is decomposed into atomic, independently verifiable claims using Claude Haiku. The decomposition prompt explicitly filters:

• Opinions and subjective assessments
• Questions and rhetorical devices
• Jokes and sarcasm
• Editorialization (loaded words separated from factual substance)

A typical tweet yields 3-8 atomic claims. Each claim proceeds through verification independently.

4.2 Semantic Precision Analysis

Before evidence gathering, each claim undergoes semantic precision analysis to identify language patterns that affect verification difficulty and confidence:

4.3 Parallel Adversarial Evidence Gathering

For each atomic claim, six evidence streams execute in parallel:

The web search stream is particularly important. For each claim, two query variants are generated:

1. Affirmative query: Searches for evidence supporting the claim as stated.
2. Counter-query: Automatically generated by flipping 17 keyword pairs (increased/decreased, approved/rejected, confirmed/denied, etc.) to actively seek contradicting evidence.

If the initial search returns fewer than 3 sources, a query reformulation step (Claude Haiku) generates 2-3 alternative search queries with different vocabulary to improve recall.

4.4 Source Classification

All gathered evidence items are classified by Claude Haiku into three stances:

• Supporting: Evidence that corroborates the claim
• Opposing: Evidence that contradicts the claim
• Neutral: Related but neither confirming nor denying

This classification happens after evidence gathering, not before, preventing confirmation bias in the search phase.

4.5 Source Enrichment

The top 3 evidence sources (by relevance) have their full page content fetched (up to 5,000 characters each) using a multi-strategy extraction pipeline: standard fetch with browser headers, Node.js native HTTPS (different TLS fingerprint), and Google Webcache fallback. Content is extracted using Mozilla Readability for clean text parsing. This enrichment ensures the synthesis model has substantive evidence, not just search result snippets.

5.1 Source Reliability

Base reliability is derived from Media Bias/Fact Check (MBFC) factual reporting ratings, covering 3,900+ domains:

5.2 Temporal Decay

For claims about current conditions (detected via temporal semantic flags), source freshness applies a decay multiplier:

5.3 Authoritative Source Overrides

Certain data sources carry inherent authority that overrides web search noise:

• Market APIs (CoinGecko, Alpha Vantage): reliability 0.95. Live price data is authoritative for financial claims.
• Government APIs (FEC, Treasury, BLS): reliability 0.80-0.90. Official records override journalistic interpretation.
• Wikipedia entity data: Confirmed entity roles override web search contradictions for identity claims.

5.4 Verdict Determination

The weighted evidence produces a verdict through the following decision logic:

5.5 Confidence Adjustments

Final confidence scores are adjusted by semantic precision flags (-15% per high-severity absolutist/causal flag), source diversity (bonus for evidence from multiple independent outlets), and evidence freshness concentration (penalty if all evidence is older than 90 days for a current-events claim).

6.1 Ownership Chain Mapping

For 484+ tracked news sources (and 5,000+ with bias ratings), the system maintains ownership chains mapping outlets to their parent companies and ultimate owners. Data is enriched from Wikidata entity relationships, SEC EDGAR corporate filings, and manual curation for complex multi-subsidiary structures.

6.2 Political Donation Integration

Using the OpenFEC API, the system retrieves employee political donation records for media parent companies. Donations are aggregated by party affiliation and presented as percentages (e.g., "Fox Corporation: 65% Republican, 32% Democrat, 3% Other"). This data appears directly in agent replies when a fact-checked tweet references a tracked news source.

6.3 Lobbying Disclosure

Senate Lobbying Disclosure Act (LDA) filings are queried for media parent companies, surfacing active lobbying relationships that may represent conflicts of interest relevant to the claims being verified.

6.4 Integration with Verification

Source provenance data is not merely displayed alongside verdicts; it informs the verification synthesis. When the Claude Sonnet synthesis model produces the final 8-category analysis, it receives source ownership and funding data as context, enabling it to flag potential conflicts of interest in its "Source Credibility" and "Bias Analysis" output categories.

7.1 Political Lean Detection

The system estimates poster political lean from multiple signals: handle keywords (partisan terms), bio content analysis, and historical tweet patterns. Lean is expressed as a direction (left/center/right) with a confidence score (0-1). This is presented transparently in agent replies as context, not as a judgment.

7.2 Accuracy History

Every fact-check creates a permanent record. When a poster is checked multiple times, their historical accuracy rate is computed (percentage of claims that verified as true across all checks). This track record is displayed in subsequent agent replies: "This poster's claims have checked out 3 of 7 times (43% accuracy)." The system gets more informative with each interaction.

8.1 Image Content Extraction

Tweets frequently contain claims embedded in images: charts, screenshots, infographics, headlines. The agent extracts content from all images in the conversation chain (up to 4 tweets, max 4 images per tweet) using Claude Haiku Vision. The extraction prompt requires both text transcription (OCR) and visual description (objects, people, scenes, chart types), ensuring that non-textual information is captured.

8.2 Cross-Reference Against Live Data

Extracted numerical claims from images (e.g., "BTC at $200K" from a trading screenshot) are automatically cross-referenced against live market APIs. This enables the system to identify manipulated screenshots, outdated charts presented as current, or deliberately misleading visualizations.

9.1 Discovery Process

Every 90 minutes, the agent executes rotating search queries across six categories: political claims, scientific assertions, economic data, breaking news, numeric claims, and technology predictions. Results are filtered for virality (minimum 1,000 likes, 200 retweets, 5,000 follower count) and factual content indicators (numbers, dates, attribution verbs, data language).

9.2 Scoring and Selection

Candidate tweets are scored by: engagement (likes + 3x retweets), account size bonuses, claim density, and quote tweet presence. The highest-scoring tweet with verifiable factual content is selected for analysis.

9.3 Delivery as Original Tweets

Proactive fact-checks are posted as original tweets (not replies) to avoid shadow-banning from unsolicited replies. This architectural decision was informed by observed Twitter/X platform behavior where unsolicited reply patterns trigger visibility reduction. The proactive budget is capped at 200 tweets per month from a total budget of 2,800.

10.1 $TRUTH Token

$TRUTH is a Token-2022 (Solana Token Extensions) asset launched on Pump.fun. Unlike standard SPL tokens, Token-2022 supports extended functionality including transfer hooks and metadata extensions.

10.2 Token-Gated Access

Holding $TRUTH provides platform access without requiring subscription payments:

10.3 Burn Mechanism

Subscription payments made in $TRUTH tokens are burned (sent to a burn wallet), creating deflationary pressure proportional to platform usage. As B2B API adoption scales, the burn rate increases with revenue.

10.4 Rate Limit Tiers

Token holdings (or paid subscriptions) determine agent interaction limits on Twitter:

10.5 Revenue Share (Planned)

As B2B API revenue scales, a revenue share mechanism for token holders is planned. This aligns holder incentives with platform growth: the more newsrooms and enterprises that license the verification API, the more value flows to token holders.

11.1 Design Principles for Error Prevention

The system incorporates multiple safeguards against the most damaging error type: false negatives (marking true claims as false).

• Asymmetric evidence thresholds: TRUE requires any supporting evidence. FALSE requires 2+ opposing sources with combined reliability above 1.0. This asymmetry ensures that search failures default to UNVERIFIABLE rather than FALSE.
• Wikipedia entity override: When Wikipedia confirms an entity claim (e.g., a person's role or title), this overrides opposing web search noise that may reference outdated information.
• Market API authority: Live price data from financial APIs (reliability 0.95) overrides web search results for financial claims, preventing outdated articles from contradicting current market conditions.
• Temporal decay: Old sources are automatically downweighted for current-condition claims, preventing historical data from overriding present reality.

11.2 Known Limitation: ETF Flow Claims

In February 2026, the system incorrectly marked a BlackRock $117M ETF outflow claim as false. The root cause was that the evidence pipeline lacked institutional flow data sources (SoSoValue, CoinGlass), relying instead on general web search which conflated on-chain Coinbase withdrawals with ETF flows. This was corrected by adding dedicated institutional flow detection and authoritative financial data sources. The incident demonstrates both the system's failure mode (domain-specific data gaps) and its correction mechanism (adding specialized evidence streams).

11.3 Continuous Improvement

Every verification result is stored permanently. The growing dataset of 5,300+ verified claims with full evidence chains enables ongoing analysis of error patterns, evidence source reliability, and verdict distribution. The system's accuracy improves as: (a) the evidence source network expands, (b) poster accuracy histories accumulate, and (c) domain-specific verification modules are added for newly identified claim categories.